Fara-7B is designed for “computer tasks” and can therefore work independently with a mouse and keyboard.
Microsoft announced its Fara-7B model in a blog post. This is Microsoft’s first AI agent-based small language model (SLM). It runs locally on PCs and is autonomous enough to perform tasks.
Small but Mighty
According to Microsoft, Fara-7B can deliver “state-of-the-art performance” compared to other models in this class, despite its 7 billion parameters. The model also competes with larger agent systems that require more resources and use multiple LLMs. Microsoft claims that Fara-7B performs better than OpenAI’s GPT-4o when it comes to web browsing.
Three videos show how Fara-7B works: it visually perceives websites and doesn’t need other models to analyze the information on screen. You can see how the model buys products online, looks up information, and measures the distance between two locations on a map. This happens slowly and the user must approve certain steps themselves, such as entering account details. Still, it’s interesting to see where AI agents are evolving.
Not yet Flawless
Microsoft clearly states that Fara-7B is not yet fully reliable. During tests, the model made errors with complex operations and sometimes followed instructions incorrectly. Hallucinations also still occur. Therefore, it may only be used in an “isolated sandbox” for now, where Microsoft can prevent misuse and data loss. The company emphasizes that the model is secured against harmful prompts.
Fara-7B appears under an MIT license on Microsoft Foundry and Hugging Face and currently only works with Magnetic-UI, Microsoft’s experimental AI platform. Later, a variant optimized for Windows 11 Copilot+ PCs will be released, which have special hardware for local AI models.